SpaceX (Trans): AI turns profitable for the first time; targeting 100bn ARR by year-end

Below is Dolphin Research's transcript of SpaceX FY26Q2 earnings call

I. Key financial takeaways

1. Revenue and ARR guide: Management expects ARR to reach $100bn by year-end, based on annualizing Dec monthly run-rate including Cursor. In the first weeks of Q3, the company signed an additional $6.7bn in cloud services contracts, spanning six months and ramping from Oct.

2. Long-term revenue target pulled forward: Pre-IPO internal targets had revenue (not ARR) reaching $1tn in 2031. This is now pulled forward to 2030, with a non-zero chance of hitting it in 2029.

3. Key metrics this quarter

Consolidated: revenue of $7.8bn (+92% YoY vs. $4.1bn a year ago); net loss of $541mn, narrowing by $467mn YoY; Adj. EBITDA of $3.5bn (+191% YoY vs. $1.2bn).Space: revenue of $962mn (+55% QoQ; +29% YoY), with costs up $389mn YoY; Adj. EBITDA loss of $205mn, driven by accelerated Starship R&D.Connectivity: revenue of $4.3bn (+32% QoQ; +66% YoY); costs up $970mn (~+58% YoY); OP rose 79% YoY to $1.7bn with OPM expanding ~300bps; Adj. EBITDA up 64% YoY to $2.6bn; enterprise and Gov. revenue +108% YoY.AI: revenue of $2.6bn (+213% QoQ; +247% YoY), with $1.6bn incremental from cloud agreements; ads +7% QoQ; costs up $1.6bn YoY; net OP loss narrowed to $1.3bn; Adj. EBITDA turned positive to $1.1bn.

4. Capex and returns: Total capex was approx. $18.4bn, of which ~$15.8bn went to AI compute infrastructure, with the remainder to Starship and launch sites, satellite production, and global ground stations. The company also paid $856mn under a spectrum credit arrangement for the EchoStar deal. Management said payback on incremental compute capex is under one year and guided capex to remain similar over the next two quarters.

5. Capital structure and balance sheet: The company completed its IPO, raising net proceeds of ~$85.7bn; then issued $25bn inaugural IG senior notes, using part of the proceeds to repay a $20bn bridge. The notes were issued in five tranches with a WAC of 5.855% and avg. tenor of 11.7 years. Quarter-end cash, cash equivalents and marketable securities were $100bn, with backlog of $47.5bn. Capital allocation remains focused on Starship, next-gen Starlink, and AI compute, while maintaining an investment-grade balance sheet.

II. Earnings call details

2.1 Management remarks

1. Starship progress and lift narrative

Two successful Starship V3 flights were completed in the past 90 days, with Flight 13 validating core capabilities for on-orbit missions and tower catch on return to Starbase. Flight 14 will be the first to place V3 Starlink comms satellites into operational orbit.The plan is to attempt both first- and second-stage recovery for Starship V3 this year, with a possible tower catch as soon as the next flight. The target is payload capacity at 4x Falcon 9 and launch cost at roughly one-tenth.Infrastructure is being built for thousands of Starship launches per year, including accelerated Raptor and airframe production, Gigabase construction, and multi-stand activation at Starbase and Cape Canaveral pads 39A and 37. Falcon currently delivers ~2,500 tons to orbit annually vs. ~300 tons combined for the rest of the world, implying 80%–90% global share for SpaceX. Starship targets well over 1m tons per year, ultimately ~10m tons.

2. Starlink broadband and consumer biz.

There were 78 launches in 1H, delivering 1,041 tons to orbit, mostly for internal Starlink missions. Q2 net adds exceeded 1.7mn consumer subs globally, a record quarter (vs. 1.4mn in Q1), with ARPU steady at $66 per month.By quarter-end, the service covered 167 markets (Elon verbally cited 170), with new markets being activated post-quarter. As of Jun 30, ~10,200 Starlink broadband and mobile satellites were on orbit, including ~9,600 broadband satellites providing ~800 Tbps of total downlink capacity.The latest Starship flight completed a V3 Starlink speedrun, achieving inter-satellite laser mesh across every link per satellite, with subsequent Starship missions to deploy V3 satellites into the constellation for service. A single V3 satellite has an order-of-magnitude higher capability than V2, and the number of V3 satellites launched will also be roughly an order of magnitude higher. Bandwidth delivery will rise by about two orders of magnitude; even if revenue per bit falls 10x, Starlink revenue could still increase 10x. Management believes Starlink could carry a majority of global internet traffic (in countries where permitted) in under 10 years.

3. Starlink enterprise and Gov.

Management expects enterprise and Gov. revenue to be at least as large as consumer, and likely larger. In Q2, the company signed a major deal with American Airlines and activated new airline partners including Southwest, Virgin Atlantic, Iberia, and Aer Lingus. Aviation is one of the largest segments with <10% penetration.One airline reported that passengers were switching to shorter, connecting itineraries to fly on Starlink-equipped flights, an unprecedented behavior in the industry. On the Gov. side, the company won over $6bn in contracts in the U.S. in Q2, supporting critical Space Force comms and sensing programs, and sees room for growth next year.

4. Starlink Mobile and spectrum

In Q2, SpaceX added mobile partnerships with SoftBank, NTT Docomo, and Spark New Zealand. Near-term priority is to deploy Mobile V2 satellites via Starship to prepare for integration of 65 MHz EchoStar spectrum later next year. The FCC has approved the spectrum transfer, viewed as a foundational competitive edge for Starlink Mobile.

5. AI compute and cloud

In Q2, the company signed compute cloud agreements for Colossus and Colossus 2 sites with customers such as Google and Anthropic. New hosted deals carry very high incremental EBITDA margins.Nameplate power capacity reached 1.4 GW at quarter-end, up from 1.0 GW in Q1 and 400 MW a year ago, and is expected to exceed 2 GW by year-end. By end-2025, cumulative online capacity will be multiples of that, closer to 10 GW than 5 GW.The compute roadmap has standardized exclusively on NVIDIA, with Rubin deemed the best architecture. STARMIND AI satellites are essentially optimized Rubin NVL72 computers and are slated to launch starting next year; the same design will also be deployed on the ground given its superior cost and performance vs. standard rack designs.

6. Grok models and Cursor acquisition

Grok 4.5 was released in Jul to positive enterprise feedback, with token consumption tripling immediately post-launch. Grok 4.6 is expected next week, 4.7 in about 3–4 weeks, and Grok 5 before year-end.Grok 5 will train on SpaceX's full corpus of 25 years of data, which management expects will make Grok the strongest engineering model. Regulatory hurdles for the Cursor acquisition have largely cleared, with closing expected soon, followed by team integration and combined sales.

2.2 Q&A

Q: How does the enterprise and Gov. backlog in Connectivity convert to revenue, and what does it imply for growth over the next two years?

A: We are very optimistic on enterprise. On Gov., we won over $6bn in the U.S. this quarter, awarded in tranches and tied to new capabilities coming online, with some still subject to competition, but underpinned by contracts across multiple phenomenologies. These include both recently announced programs and new ones to come.Enterprise revenue is sticky and we have not lost a single enterprise customer. Customers are very satisfied with current Starlink capabilities. Aviation penetration is only ~10%, implying significant runway. At sea, Starlink materially expands TAM: many vessels have little to no connectivity because they cannot afford VSAT, and they can retrofit directly with Starlink Maritime.One more point: to win large enterprise deals, we must prove extremely high reliability and uptime. Many enterprises still retain early impressions of intermittent uptime from Starlink's early days, so we are building a sizable enterprise sales force to visit large enterprise and Gov. customers and demonstrate today's uptime and low latency. The goal is to position Starlink as a primary link, not backup. This requires one-on-one education on the system's current state and roadmap. We expect enterprise revenue to significantly exceed consumer.

Q: On the 2027 nameplate capacity closer to 10 GW than 5 GW, how certain is visibility on power permits, chips, and turbines?

A: On power, cooling, and electrical gear, our target actually far exceeds that GW number. A provisional goal is to bring 20 GW of power and cooling online by end-2025. We likely will not hit 20 GW, but we aim to stack a portfolio of projects totaling 20 GW; with some delays, say a quarter slipping, we still expect to be near 15 GW at the plant level.The goal is to have power, cooling, and electrical inventory well in excess of GPUs, which is rational given GPUs' cost share vs. the rest of the system. SpaceX is exceptionally strong in hardware. Ex-China, SpaceX and Tesla may be the strongest hardware companies on Earth—possibly including China as well. By redeploying even a small slice of our rocket and satellite hardware capability into scaled data centers, we are already seeing significant gains. Our consensus with NVIDIA is that next year we will receive a very high share of its GPUs.

Q: You previously said robots were the most underestimated part of the SpaceX story. Why, and how do robots expand the TAM?

A: The more precise point is that the market underestimates future connectivity demand. Robots—especially AI-enabled robots, from cars to humanoids—will dramatically increase the need for connectivity.Satellites themselves are robots: autonomous and maintenance-free. Looking ahead, and sooner than people think, robots will scale manufacturing on the Moon. That may sound like sci-fi, but it will happen, enabling a lunar mass driver; with lunar mass drivers plus lunar solar and radiators, we could scale intelligence launched to space to 1,000x the Earth economy, perhaps even a million-fold. We will ship large tonnage to the Moon to build factories, where robots will be essential.

Q: Capex outlook by segment? Where does 'buy' beat 'build' to accelerate growth (e.g., mobile)? What are the next capital needs for deep vertical integration?

A: For the rest of this year, assume capex in the next two quarters is similar to this quarter. Longer term, capex differs by nature: in AI compute, we can deploy capital with sub-one-year payback, making it almost a COGS-like item given the speed of monetization and revenue conversion. This is unlike SpaceX's historical capex, which built assets that pay back over decades, such as launch pads. We manage capital strictly by ROI.Two adders: first, we expect ARR above $100bn in Dec. Second, our internal forecast for $1tn revenue (revenue, not ARR) moved up from 2031 to 2030. That was the pre-IPO view for 2031; now we expect 2030, with a non-zero chance in 2029.

Q: What did Flight 13 teach you on Starship; what beat or trailed expectations? Implications for rapid reuse and tiles?

A: We were a bit concerned about Flight 13—Apollo 13 lore—but the flight went extremely well. Heat shield tiles look very robust; the ship is still floating and we expect to recover it for analysis. Overall results were strong, so pending regulatory approval, we will attempt a tower catch on the next flight, tentatively later this month.We expect flight cadence to ramp quickly, targeting at least daily launches in about a year, if not more.

Q: Which rocket and satellite engineering capabilities unlocked acceleration in compute deployment?

A: 'Rocket science' is a synonym for 'extremely hard'—and that is our daily job. Every launch vehicle is trying very hard to blow itself up every flight; the engineering struggle is to persuade the rocket not to do that and deliver payload to orbit. Competitors' rockets do blow up sometimes—this is not an indictment of their abilities; they are smart teams, but rockets still blow up.So shifting even a small slice of the capability required to make a giant, reusable rocket fly reliably at high cadence into ground data centers is, frankly, like the Yankees playing in a Little League. Compared with reusable heavy-lift at high frequency, ground data centers are mundane problems. Starlink satellites are also alien-grade tech; nothing matches V2 today, and V3 is over an order of magnitude better than V2. Applying even a small portion of that engineering to data centers yields excellent results.

Q: How will compute supply-demand evolve, and can premium pricing persist?

A: Best guess for monetization is roughly $30–$50 per watt; this is just a guess. AI capability per watt is improving quickly, so usefulness per watt is rising fast. Smaller models are doing astonishing things; there is even a case that H100-class could provide AI as useful as human engineers, which is not impossible on current trends. If that happens, utility is enormous. Think about AI 12 months ago, or two years ago—two years ago looks museum-grade.Given the current pace, by end-2025 it is unclear what AI cannot do. On logic and storage output, the bottleneck today is storage. Storage output grows ~20% a year, which is very fast for any large, mature industry. But demand is not growing 20%—it is 200%, perhaps higher. Demand growth far outstrips supply growth, so basic economics says prices should rise, not fall.

Q: How is the $100bn ARR decomposed, and what gets you there?

A: All three segments are growing, but AI cloud is the largest driver. We booked additional new business in the first two weeks of Jul. Previously announced partnerships with Google and Anthropic will ramp later this quarter or in Oct, providing strong momentum.Our own models are performing well, and Cursor will be consolidated. All three segments contribute, with cloud as the largest piece.

Q: What gives you confidence to pull the revenue target forward a year?

A: To be clear, $100bn ARR in Dec is not a question mark; that is the 'do nothing else' baseline, so actual could be higher, and likely will be. The mix of revenue sources should resemble what we shared pre-IPO, just shifted earlier by a year.Also, V3 satellites will accelerate Starlink comms revenue. V3 has >10x V2 capability, and we will launch ~10x as many, while direct-to-cell is deploying. Some observers still extrapolate V2-era curves, but V3 is a step-change in magnitude.

Q: When does Starlink Mobile start launching V2 satellites? Appetite for additional spectrum? How big is mobile vs. broadband?

A: Nomenclature can be confusing: 'Mobile V2' refers to the second-gen mobile satellites, while V3 refers to next-gen broadband satellites. Mobile V2 launches begin next year.The current system uses only ~5 MHz via local carrier partnerships. With EchoStar spectrum, we will have 65 MHz available, a major capability step-up; satellite count will also be ~10x. In simplified terms, next-gen satellites plus EchoStar spectrum will make Starlink Mobile ~100x better (10x times 10x). Even today, the service supports voice and video calls via Signal and WhatsApp.We expect strong user adoption. Top-down, the three major U.S. carriers—AT&T, Verizon, and T-Mobile—generate roughly $600bn annually, and we expect to take a meaningful share by offering better service: eliminating dead zones and performing better during natural disasters, since space is actually a very calm environment despite Hollywood depictions. Satellites start launching next year, with service expected by late next year.

Q: Of the added compute through 2027, how much for internal Grok training vs. external rental?

A: Over time, a smaller share will be dedicated to Grok training, with a larger share for inference or external training and inference rental. Roughly, Grok training will consume about 10% of capacity.

Q: Becoming a true fourth U.S. operator implies >$100bn long-term mobile capex (mostly spectrum and fresh infra). How do you weigh build vs. buy, and what is the base case?

A: The EchoStar spectrum acquired carries terrestrial usage rights, and we do intend to build the ground portion. Future capacity will thus come from both satellites and the terrestrial hardware and systems needed for a true mobile service. We are not disclosing capex now—not due to immaturity, but by choice.One hint: we can mount cellular radios on mounts that already hold Starlink broadband antennas, deploying small femto cells nationwide as needed. This avoids a pre-launch, upfront multibillion-dollar spend on low-band spectrum and macro sites. We have new ideas on deployment, so it is premature to guide capex, but we are confident it will be highly capital-efficient.This is important: we do not need extremely expensive, hard-to-site macro towers. We are confident we can deploy a large number of small cells—antennas that also provide connectivity in mobile bands—nearly everywhere. Installed on residential and commercial roofs with clear views, they can connect directly to handsets, and we believe bandwidth and quality may exceed incumbent cellular providers.

Q: Updates on the Starship HLS (human landing system)?

A: Before flying crew, Starship must achieve very high reliability on satellite launches. We expect high cadence to reach crew-safety levels quickly, likely by late next year.Milestones: on-orbit propellant transfer is crucial for both SpaceX internal goals and HLS. Artemis III is slated for next year, with docking to Orion; then an uncrewed direct lunar cargo mission; and the target is a landing in 2028.

Q: Post-close product roadmap for Cursor integration?

A: We will disclose outside the earnings call. We prefer not to pre-announce before regulatory clearance; closing is near, and we are managing the process carefully.

Q: After V3 broadband satellites begin launching, how quickly will Starlink service improve?

A: We need to reach a V3 critical mass, at least on the order of 1,000 satellites. Timing is roughly Q2 next year.

Q: Are heat shield tiles the last hard problem for Starship?

A: We prefer not to overstate, but tiles can be considered solved at this point. We will still inspect the floating ship carefully, but data and visual checks indicate tiles—the single largest issue—are resolved.This does not mean improvements stop. But on full and rapid reuse, we see no technical blockers.

Risk disclosure and statements:Dolphin Research Disclaimer and General Disclosure